- 學位論文
基於熱圖像識別技術的實時雨水檢測系統
Real-Time Rain Detection System Based on Thermal Image Recognition Technology
摘要
現代社會繁忙，在建築物內工作的民眾不一定能得知室外天氣的實時狀態，為了能迅速理解室外大概的天氣狀態，民眾通常會使用網路或APP來查詢，查詢到的結果通常是大區域地區的預估狀況，但就小區域而言此資訊不具有實時性。 本論文提出基於熱圖像識別技術的實時雨水檢測系統，紅外線熱像儀搭配樹莓派3B+偵測發熱的陶瓷加熱燈(Ceramic Heat Lamp)當作辨識目標物。每秒讀取一次陶瓷加熱燈當下狀態，先將它加熱到180度，水一碰到加熱面就會立刻蒸發，此加熱狀態被定義為Hot；當水接觸到高溫的陶瓷加熱燈時蒸發後，加熱面接觸水的部分會產生溫度變化，此狀態被定義為Rain。當系統斷電陶瓷加熱燈未加熱或是未加熱到180度溫度不足，此狀態被定義為Others。 在實驗中，將Hot、Rain和Others三類熱影像圖片集資料放入Google Teachable Machine平台使用機器學習訓練，Hot 150張、Rain 150張、Others 300張做50回訓練，每批16份資料，學習率定為0.001，經過訓練後生成模型和標籤後放入樹莓派。為了增加辨識處理效率，在樹莓派上安裝神經運算棒二代，接著建立OpenVINO環境，安裝OpenCV和TensorFlow Lite為物件辨識和邊緣運算使用，讀取訓練過的模型和標籤，便可開始執行辨識，讀取熱像儀的鏡頭畫面，會立刻顯示信心度，經測試Hot信心度能達到平均99.088%，Rain信心度能達到平均98.792%。 未來，提出的商業化系統可以為當地區域提供可靠和實時的雨量檢測系統，其中包括百貨公司、地鐵站、地下商場和其他設施。
並列摘要
In the busy modern society, people working in buildings may not know the real-time status of outdoor weather. To immediately understand the outdoor weather status, people usually use the Internet or APP to obtain it, and the weather information is usually the status of a large region. However, the information is not immediate for a local area. This paper proposes the real-time rain detection system based on thermal image recognition technology, which includes an infrared thermal imaging camera, the development board of Raspberry Pi 3B+, and a ceramic heat lamp as the detection target. The system reads the current state of the ceramic heat lamp once per second and heats it to 180 degrees. When rainwater drops on the ceramic heat lamp, it is evaporated quickly. The 180-degree heat state names Hot state. When rainwater contacts the high-temperature ceramic heat lamp, the part of its surface will lead to a reduced temperature change, whose form names Rain state. When the system is powered off, turned off, or not heated to 180 degrees far away, this state names Others state. In the experiments, we put the three types of thermal image data sets of Hot, Rain, and Others into the Google Teachable Machine (TM) platform and used machine learning to train the thermal images. We utilized a Hot data set of 150 pictures, a Rain data set of 150 photos, and Others data set of 300 ones to train the model with 50 epochs, a batch size of 16 pictures, and a learning rate of 0.001. After training finish, the TM platform generated the model and label file, which was implemented into the development board of Raspberry Pi. To increase the efficiency of detection processing, the development board of the Raspberry Pi installed Intel Movidius Neural Compute Stick 2, then created the OpenVINO environment, installed OpenCV and TensorFlow Lite for object detection based on edge computing. The system read the trained model and label file and then started to detect. By capturing the photos via the FLIR-C3 thermal camera will immediately obtain the confidence coefficient. After testing, the confidence of the Hot state had an average of 99.088%, and that of the Rain state was 98.792%. In the future, the proposed commercialized systems can provide a reliable and real-time rain detection system to a local area, where includes department stores, MRT stations, underground shopping malls, and other facilities.